Machine learning in finance

• Fraud detection
• Credit risk analysis
• Return prediction
• Valuation
• Text analysis
• Time series forecasting

Regression vs Classification

• Regression means to predict a continuous variable (not necessarily linear regression).
• Classification is to predict a categorical variable. Binary or multiclass.

Underfitting and Overfitting

• Fitting or training a model means estimating its parameters (like linear regression).

Model complexity and shrinkage

• We can create more complex models by increasing the number of parameters (like adding variables in a linear regression).

Train and Test

• To check that we have not overfit our model, we train and test.
• Split data randomly into train and test samples. Train (fit) on the training data. Test on test data.
  • Test data is also called holdout data.
  • Also called out-of-sample testing.
• Performance on the test data is the performance that we can expect on new data.

Test criteria

• How do we decide if performance is good or bad?
• For continuous variables,
  • usually want to achieve a low sum of squared errors
  • equivalently, achieve a high \(R^2\).

\[R^2 = 1 - \frac{\sum (y_i - \hat y_i)^2}{\sum (y_i - \bar{y})^2}\]

• For categorical, can use % accurately classified.

Validation

• We could consider different models - for example, models of different complexity - and train them all, test them all on the test data, and then choose the model that performs best on the test data.

Data

• Download irrelevant_features.xlsx from the course website
• Upload it to Julius and ask Julius to read it and describe it.

The data was created by generating 51 sets of 100 standard normals.

• The first 50 sets are labeled x1, …, x50.
  
• The last set was used as the noise to generate y1 as x1 + noise.
• So, x2, …, x50 are irrelevant for y1, but they may be correlated with y1 by chance.

Linear regression

• Ask Julius to do a train-test split of the data with 20% of the data in the test set.
• Ask Julius to train a linear regressor on the training data with x1, …, x50 as the features and y1 as the target.
• Ask Julius to report the coefficient estimates.
• Ask Julius to compute the R-squared on the test data.

Shrinkage

• To induce selection of parameter values that are less influential, we can penalize large values.
• Usually in linear regression, we minimize SSE (sum of squared errors).
• LASSO: minimize SSE + penalty \(\times\) sum of \(|\beta_i|\).
• Ridge: minimize SSE + penalty \(\times\) sum of \(\beta_i^2\).

Train and Test

• Ask Julius to train a ridge regressor on the training data, report the parameter estimates, and compute the R-squared on the test data.
• Ask Julius to traing a lasso regressor on the training data, report the parameter estimates, and compute the R-squared on the test data.
• The penalty in lasso and ridge is called alpha in scikit-learn. You can specify it.

Optimizing the penalty

• The penalty is called a hyperparameter, because it is not fit by training but instead is specified in advance.
• The penalty controls the degree of complexity. Larger penalty = less complex.
• To optimize the penalty, we can cross validate.

Cross Validation

• Split the training data into random subsets, say, \(A, B, C, D\), and \(E\).
• Use \(A \cup B \cup C \cup D\) as training data and test on \(E\).
  
• Then use \(B \cup C \cup D \cup E\) as training data and test on \(A\).
  
• Then, …, until we have trained and tested 5 times.

• Average the 5 test scores.
• Repeat for each model configuration (each hyperparameter value).
• Choose the hyperparameter value with the highest average score.
• Then proceed to testing on the test data.

Cross Validate Lasso and Ridge

• Ask Julius to run GridSearchCV on lasso regression.
• You can specify a set of alpha values to try.
• Ask Julius what the optimal alpha (penalty) is.
• Ask Julius what the coefficient estimates are.
• Ask Julius what the score is on the test data.
• Repeat for ridge regression.